YR Sports Lab
Who Advances at the 2026 World Cup?
Title Odds from 5 Statistical Models × 600,000 Monte Carlo Simulations
The first ever 48-team, 12-group, 104-match World Cup is under way in North America. This article feeds five data sources, official FIFA ranking points, Elo ratings, full-squad Transfermarkt values, historical World Cup records and recent major-tournament form, into five independent statistical models. Each model simulates the entire tournament 100,000 times, and an equal-weight ensemble pools the results. Who climbs out of the group of death? Where does the trophy end up? Everything below is answered with validated data.
Ensemble title odds: Argentina, France and Spain lead a three-way race
The win, draw and loss probabilities of all five models are averaged with equal weight for every fixture, then the full tournament is simulated 100,000 more times to produce the final title odds. Defending champions Argentina lead on 18.8%, the most expensive squad in the tournament, France, sit on 17.3%, and world number one on Elo, Spain, take third on 16.4%. Between them the top three hold more than half of all title equity. The coloured dots on each bar mark the estimate from each of the five base models: the wider the dots are spread, the more the data sources disagree about that team.
Five models, five different World Cups
Same tournament, wildly different scripts depending on the data source. The Elo model is infatuated with Spain (40.5%, with an Elo of 2157, the highest in the field), the squad-value model backs France and England (two rosters worth €2.84 billion combined), and the history model puts Brazil and Germany back on the throne (nine trophies between them, pure World Cup DNA). The Spearman rank correlation between the five models' title rankings falls as low as 0.78, and that disagreement is precisely why an ensemble is used: no single data source has a monopoly on the truth.
Why an equal-weight ensemble instead of just picking the best model?
The classic result in the forecast-combination literature is that a simple average of several reasonable models is extremely hard for any single model to beat consistently over time (Clemen, 1989). The five models here are biased in different directions: Elo overrates teams on recent winning streaks, squad value overrates star concentration, history overrates traditional powers, and FIFA points are flattened by the ceiling effects of the scoring formula. Equal weighting lets those biases cancel, and the ensemble output sits squarely inside the consensus range of five external expert models.
Advancement odds for all 12 groups: who survives the group stage
The new format splits 48 teams into 12 groups. The top two in each go straight to the round of 32, and the eight best of the 12 third-placed teams also advance, an overall qualification rate of 66.7%, the most forgiving in World Cup history. It also turns "playing for third" into a genuine tactical option. From dark to light, the stacked bars show the probability of winning the group, finishing second and advancing as a best third-placed team; the number on the right is total advancement probability. The media consensus group of death is Group I (France, Norway, Senegal, Iraq): three teams above 1850 Elo in one group, with Norway carrying Haaland (€200m valuation) yet only a 74% chance of getting through.
The most likely final
These are the final match-ups that come up most often across 100,000 ensemble simulations. Spain vs Argentina (7.0%) and France vs Argentina (6.2%) are the two dominant themes: on the bracket, Argentina (Group J) sit in a different half from Spain (Group H) and France (Group I), so the big three can only meet in the final at MetLife Stadium on 19 July. Note the long tail of the distribution: the top 12 pairings together account for only about 40% of outcomes, which means in 60% of parallel universes the final is one of more than 1,100 other match-ups. That is the chaotic nature of single-elimination football.
Roll the dice yourself: interactive Monte Carlo simulator
The simulator below ports exactly the same ensemble engine used throughout this article into your browser: it draws a scoreline for each of the 72 group matches, ranks teams on points and goal difference, picks the eight best third-placed teams, and plays out all 31 knockout matches on the official FIFA bracket. Draw a single "parallel universe" to see one complete tournament, or run ten thousand of them and watch the probabilities converge on the numbers in this article.
2026 World Cup simulation engine
"Draw one parallel universe" plays a full tournament and shows every group table and the knockout path; "Run 10,000 times" tallies title counts live so you can watch Monte Carlo convergence for yourself.
The engine shares its inputs with the article: ensemble probability = equal-weight average of the five models; group ranking by points, then goal difference, then goals scored, then head-to-head; third-place slots matched back to the group pools published by FIFA. The browser uses an unseeded generator, so every run differs.
The pre-tournament arms race: switched allegiances and new recruits
The build-up to this tournament saw a clear wave of nationality switches, with several European-academy players changing shirts. Those moves are already priced into the Transfermarkt squad values dated 4 June, which means the value model naturally counts recently recruited players in its ratings. Here are the verified headline cases:
Method, assumptions and validation: where every number comes from
1. Data and model setup
All inputs were collected and cross-validated before kick-off on 2026-06-11. The 12 groups were checked against two independent sources, Wikipedia and NBC Sports, and matched. The knockout bracket (including the group pools that feed the eight best third-placed slots) was checked against Wikipedia and ESPN, and matched. FIFA points come from the official release of 11 June, the last one before the tournament. Elo comes from live eloratings.net data reconciled with a Wikipedia snapshot. Squad values come from the total value of the official 26-player lists published by Transfermarkt (€17.25 billion across all 48 teams).
| Model | Strength source | Mapping | Calibration quality |
|---|---|---|---|
| M1 Elo | eloratings.net (2026-06-11) | Used directly, win expectancy We = 1/(1+10^(−Δr/400)) | Native scale |
| M2 FIFA points | Official FIFA ranking (2026-06-11) | FIFA formula divisor 600, equivalent conversion rating = pts × 2/3 | Official formula |
| M3 Squad value | Transfermarkt 26-player list (2026-06-04) | OLS: Elo = a + b·ln(value €M) | R² = 0.667 |
| M4 History × recent form | Historical World Cup records + recent majors | S = 3×titles + 2×finals + 1.5×semis + 0.5×wins + 0.25×appearances + 1.5×recent form, OLS: Elo = a + b·√S | R² = 0.525 |
| M5 Poisson | Composite strength from centred M1–M4 average | Double-Poisson scoreline simulation: total goals 2.65, μ = Δr/170 | Scoreline level |
| M6 Ensemble | M1–M5 | Equal-weight average of win/draw/loss probabilities per match | Combined forecast |
Shared assumptions: home advantage grants the United States, Mexico and Canada +80 Elo-equivalent points (the football literature puts home advantage at roughly 100 points, but the three hosts do not play every match at home, so a conservative value is used; the Goldman Sachs and Reade models also build in a host adjustment). Draw probability follows pD = 0.285·exp(−(Δr/700)²), giving a 28.5% draw rate between evenly matched sides, consistent with the 24–31% observed in group stages of the last three World Cups. Knockout matches tied after 90 minutes are resolved by relative win probability excluding draws, standing in for extra time and penalties. Group ranking follows points, goal difference, goals scored, head-to-head result, then a coin flip. Third-place slots are matched back to the per-match group pools published by FIFA (FIFA actually uses a 495-scenario lookup table; the two are equivalent).
2. Internal consistency and convergence checks
In each model's 100,000 simulations, title probabilities sum to exactly 100%, final berths to 200% and round-of-32 berths to 3,200%, all verified programmatically. The Monte Carlo standard error at p = 10% is roughly ±0.09 percentage points, meaning the sampling error on every probability quoted here is under 0.1 points, far smaller than the systematic differences between models. You can verify convergence yourself with the simulator above: a run of ten thousand tournaments lands within ±0.6 percentage points of the figures in this article.
3. External validation: the betting market
Betting odds are the wisdom-of-crowds benchmark for any forecast. Converting the 10–11 June outright title odds from BetMGM, DraftKings and FanDuel into implied probabilities, averaging them and normalising away the overround gives, against the ensemble: Spearman rank correlation 0.81, mean absolute error 2.2 percentage points. The biggest disagreements are Argentina (model 18.8% vs market roughly 11%) and England (model 8.6% vs market roughly 14%). The market's England premium and defending-champion discount map neatly onto the "champions' curse" discussed in the Goldman Sachs report and England's perennial popularity tax in the odds, while the academic models (Reade has Argentina at 23%, KU Leuven at 16.8%) sit closer to this article.
4. External validation: five public expert models
Five publicly verifiable models released title probabilities before the tournament. The ensemble here correlates with four of them at a Spearman rank correlation of 0.6–0.9: closest to Reade's Poisson regression (0.90), then KU Leuven's Elo model (0.80) and the Opta supercomputer (0.76). The Stern model published only a top four (a rank correlation on n = 4 is statistically meaningless and is listed for reference only). The title pick of all five external models falls inside this article's own top five.
5. Sensitivity analysis and limitations
Home advantage is the assumption with the largest influence. Re-running 50,000 simulations with +80 replaced by 0 or +160 moves Mexico's title probability across 0.5%, 2.8% and 9.8%, and the United States across 0.3%, 1.8% and 7.0%, but the order of the top three (Argentina, France, Spain) is unchanged under all three settings, with title probabilities shifting only 1–3 percentage points. In other words, the host assumption dominates the numbers for the three host nations but does not disturb the overall conclusion.